huggingface/speech-to-speech — reverse-engineered prompt

Build me a Python speech to speech voice agent that I can run locally with open source models. I want to talk into my mic, have it detect when I’m speaking, turn speech into text, send that text to a chat model, and then speak the reply back out loud with low latency. Please make it modular so the speech recognition, language model, and voice output pieces can be swapped, but give me a sensible default setup that works out of the box.

I’d like a simple CLI for the main workflows. It should support a normal local mode, a server mode where the models run on one machine and audio streams from a client, and a websocket or realtime style mode so another app can connect to it. A Mac friendly local path would be great, and it should also be usable on Linux. Please include Docker support, basic tests, and clear setup and usage docs.

If current model or API details have changed, look up the latest docs and wire in the closest working version.